Embedding Model for Missing EHR Data Modalities
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing healthcare technologies face challenges in predicting patient outcomes due to missing and incomplete data from electronic health records (EHRs), which are multi-modal and non-standardized, limiting the ability to accurately forecast patient discharge and resource allocation in healthcare settings.
Innovation Solution
A method and system that utilize an embedding model to generate complete embeddings for patient data by combining available and missing data modalities through an episode snapshot graph, allowing for accurate prediction of patient outcomes such as length of stay and discharge destination using multi-modal data from EHRs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If multi-modal data from EHRs is used for prediction, then the comprehensiveness of patient information is improved, but the complexity of data processing increases due to missing and non-standardized data modalities
Solution Approach 1:
The patent introduces an embedding model as an intermediary that transforms multi-modal patient data (labs, vitals, demographics, notes) into a unified continuous vector representation. This embedding layer mediates between the heterogeneous data sources and the prediction model, handling missing modalities by learning from available data patterns without requiring complete standardized inputs from all patients.
Solution Approach 2:
The system changes the parameter representation of patient data by converting discrete, categorical, and heterogeneous medical data into continuous embedding vectors. This parameter transformation allows the model to process incomplete and non-standardized data by learning meaningful representations from the available modalities, effectively changing how data completeness is measured and utilized.
2Ease of manufacture
If traditional statistical and machine learning techniques are used, then the simplicity of implementation is maintained, but the ability to extract useful medical insights from incomplete EHR data deteriorates
Solution Approach 1:
The patent replaces traditional statistical and machine learning approaches with a deep learning architecture that uses embedding models. This substitution enables the system to automatically learn from incomplete and heterogeneous data patterns, achieving superior prediction accuracy for patient outcomes while maintaining implementation feasibility through modular model design.
3Measurement precision
If complete patient data is required for accurate prediction, then the prediction accuracy is improved, but the applicability to real-world healthcare settings deteriorates due to missing data modalities
Solution Approach 1:
The system implements partial action by effectively utilizing only the data modalities that are available for each patient, rather than requiring complete data from all modalities. The embedding model learns to generate accurate predictions by processing subsets of available data (e.g., labs without vitals, or demographics without clinical notes), making the system adaptable to real-world incomplete data scenarios while maintaining prediction accuracy.
Data Source
AI summary
A method for predicting a patient outcome from a caretaker episode includes receiving a current episode snapshot of the caretaker episode comprising multi-modal data of the patient from an electronic health records (EHR) system, the multi-modal data including one or more available data modalities and one or more missing data modalities. The multi-modal data is applied as input to an embedding model having a submodel for each of the data modalities. A first embedding is generated for each of the available data modalities. A second embedding is generated for each of the missing data modalities using corresponding embeddings of neighbors in an episode snapshot graph. The first and second embeddings are combined to obtain a complete embedding. The patient outcome is predicted based on the complete embedding for the current episode snapshot using a machine learning component which has been trained using patient outcomes of the historical episode snapshots.